---
title: internlm-xcomposer2d5-7b-reward
canonical_url: "https://www.modelscope.cn/models/Shanghai_AI_Laboratory/internlm-xcomposer2d5-7b-reward"
md_url: "https://www.modelscope.cn/models/Shanghai_AI_Laboratory/internlm-xcomposer2d5-7b-reward.md"
repository: Shanghai_AI_Laboratory/internlm-xcomposer2d5-7b-reward
last_updated: 2025-05-20
license: other
pipeline_tag: any-to-any
tasks:
- any-to-any
model_type:
- internlm2
architectures:
- InternLMXComposer2ForCausalLM
library_name:
- pytorch
- transformer
frameworks:
- pytorch
language:
- en
- zh
inference_backends:
- "deploy_task vlm"
- "lmdeploy_turbomind 0.9.1"
downloads: 1228
stars: 0
tags:
- "reward model"
---
# internlm-xcomposer2d5-7b-reward
> internlm-xcomposer2d5-7b-reward - Shanghai_AI_Laboratory 在 ModelScope 开源的模型。InternLM-XComposer-2.5-Reward
Shanghai_AI_Laboratory/internlm-xcomposer2d5-7b-reward 是 ModelScope 魔搭社区上的any-to-any模型，采用 other 许可，可用 deploy_task vlm、lmdeploy_turbomind 0.9.1 部署。
- **Repository**: Shanghai_AI_Laboratory/internlm-xcomposer2d5-7b-reward
- **License**: other
- **Tasks**: any-to-any
- **Inference backends**: deploy_task vlm, lmdeploy_turbomind 0.9.1
- **Tags**: reward model
- **Downloads**: 1228
- **Stars**: 0
- **Last updated**: 2025-05-20
Source: https://www.modelscope.cn/models/Shanghai_AI_Laboratory/internlm-xcomposer2d5-7b-reward
---
InternLM-XComposer-2.5-Reward
[💻Github Repo](https://github.com/InternLM/InternLM-XComposer)
[Paper (ACL 2025 Findings)](https://huggingface.co/papers/2501.12368)
## News and Updates
- `2025.05` 🎉🎉🎉 [InternLM-XComposer2.5-Reward](https://huggingface.co/internlm/internlm-xcomposer2d5-7b-reward) was accepted to [ACL 2025 Findings](https://2025.aclweb.org/). See you in Vienna!
- `2025.05` 🎉🎉🎉 We release the [training code](https://github.com/InternLM/InternLM-XComposer/tree/main/InternLM-XComposer-2.5-Reward/training) and [a portion of our training data](https://huggingface.co/datasets/ChrisDing1105/MMIF-23k) of [InternLM-XComposer2.5-7B-Reward](https://huggingface.co/internlm/internlm-xcomposer2d5-7b-reward).
- `2025.04` 🎉🎉🎉 We release the [evaluation code](https://github.com/InternLM/InternLM-XComposer/tree/main/InternLM-XComposer-2.5-Reward/evaluation) of [InternLM-XComposer2.5-7B-Reward](https://huggingface.co/internlm/internlm-xcomposer2d5-7b-reward), including [VL-RewardBench](https://huggingface.co/datasets/MMInstruction/VL-RewardBench), [RewardBench](https://huggingface.co/datasets/allenai/reward-bench), and [RM-Bench](https://github.com/THU-KEG/RM-Bench)
### Introduction
**InternLM-XComposer2.5-Reward** is a multi-modal reward model trained on the foundation of [internlm/internlm-xcomposer2d5-7b](https://huggingface.co/internlm/internlm-xcomposer2d5-7b). This model has been trained using preference samples across text, image and video domains, and assigning appropriate reward scores that align with human preferences.
## Performance Evaluation
- Result on [VLRewardBench](https://arxiv.org/abs/2411.17451)
| Models | General | Hallucination | Reasoning | Overall | Macro |
| --- | --- | --- | --- | --- | --- |
| InternLM-XComposer2.5-7B-Reward | 84.7 | 62.5 | 62.9 | 65.8 | 70.0 |
- Result on [RewardBench](https://arxiv.org/abs/2403.13787)
| Models | Score | Chat | Chat Hard | Safety | Reasoning |
| --- | --- | --- | --- | --- | --- |
| InternLM-XComposer2.5-7B-Reward | 88.6 | 90.8 | 83.8 | 87.8 | 90.0 |
- Result on [RM-Bench](https://arxiv.org/abs/2410.16184)
| Models | Chat | Math | Code | Safety | Easy | Normal | Hard | Average |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| InternLM-XComposer2.5-7B-Reward | 65.5 | 55.9 | 51.7 | 93.8 | 87.5 | 71.3 | 47.4 | 68.8 |
### Basic Usage
Here is an example of how to use the model to get the reward score of a chat, compare two chats, or rank multiple chats.
```python
import torch
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained(
"internlm/internlm-xcomposer2d5-7b-reward",
device_map="cuda",
torch_dtype=torch.float16,
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("internlm/internlm-xcomposer2d5-7b-reward", trust_remote_code=True)
model.tokenizer = tokenizer
chat_1 = [
{"role": "user", "content": 'I want to buy a car from the input image, analyze the advantages and weaknesses.'},
{"role": "assistant", "content": "The car in the image is a Mercedes-Benz G-Class. This is a luxury SUV known for its ruggedness, off-road capability, and iconic boxy design. Here's an analysis of its advantages and weaknesses: Advantages: Off-road capability: The G-Class is built on a robust ladder frame chassis and has a powerful engine, making it highly capable in off-road conditions. It features a low-range transfer case, locking differentials, and a high ground clearance. Luxury and comfort: Despite its rugged exterior, the G-Class offers a luxurious interior with high-quality materials, comfortable seats, and advanced technology features. Iconic design: The distinctive boxy shape of the G-Class has made it a cult classic. It's instantly recognizable and has a timeless appeal. Strong resale value: Due to its exclusivity and desirability, the G-Class tends to hold its value well over time. Weaknesses: Fuel efficiency: The G-Class is not known for its fuel efficiency, especially with the larger engine options. Ride quality: The firm suspension is designed for off-road performance, which can make the ride a bit bumpy on paved roads. Price: The G-Class is a very expensive vehicle, even for a luxury SUV. Size and maneuverability: The large size of the G-Class can make it challenging to park and maneuver in tight spaces."}
]
chat_2 = [
{"role": "user", "content": 'I want to buy a car from the input image, analyze the advantages and weaknesses.'},
{"role": "assistant", "content": "Based on the image, it appears to be a Ferrari F8 Tributo. Advantages: Performance: Ferraris are renowned for their exceptional performance. The F8 Tributo is no exception, boasting a powerful twin-turbocharged V8 engine that delivers exhilarating acceleration and top speed. Handling: Ferraris are known for their precise handling and agility. The F8 Tributo continues this tradition, with advanced aerodynamics and sophisticated suspension systems ensuring a thrilling driving experience. Design: Ferrari is synonymous with iconic and head-turning designs. The F8 Tributo is a testament to this, featuring sleek lines, aggressive styling, and a captivating presence. Exclusivity: Owning a Ferrari is a status symbol. It represents luxury, exclusivity, and a passion for high-performance automobiles. Weaknesses: Price: Ferraris come with a hefty price tag. The F8 Tributo is no exception, making it an investment for those with significant financial resources. Fuel Efficiency: High-performance sports cars like the F8 Tributo are not known for their fuel efficiency. You can expect lower miles per gallon compared to everyday vehicles. Maintenance Costs: Owning a Ferrari comes with associated maintenance costs, which can be higher than those for regular cars. Practicality: The F8 Tributo is primarily a two-seater sports car, making it less practical for everyday use or carrying passengers. Ride Comfort: While the F8 Tributo offers a thrilling driving experience, its stiff suspension might not be ideal for long-distance comfort."}
]
image = ['./examples/cars1.jpg']
hd_num = 9
# get reward score for a single chat
with torch.autocast(device_type='cuda', dtype=torch.float16):
score1 = model.get_score(chat_1, image, hd_num=hd_num)
score2 = model.get_score(chat_2, image, hd_num=hd_num)
print("score1: ", score1)
print("score2: ", score2)
# >>> score1: 5.76
# >>> score2: -2.84375
# batch inference, get multiple scores at once
with torch.autocast(device_type='cuda', dtype=torch.float16):
scores = model.get_scores([chat_1, chat_2], [image, image], hd_num=hd_num)
print("scores: ", scores)
# >>> scores: [5.76171875, -2.845703125]
# compare whether chat_1 is better than chat_2
with torch.autocast(device_type='cuda', dtype=torch.float16):
compare_res = model.compare(chat_1, image, chat_2, image, hd_num=hd_num)
print("compare_res: ", compare_res)
# >>> compare_res: True
# rank multiple chats, it will return the ranking index of each chat
# the chat with the highest score will have ranking index as 0
with torch.autocast(device_type='cuda', dtype=torch.float16):
rank_res = model.rank([chat_1, chat_2], [image, image], hd_num=hd_num)
print("rank_res: ", rank_res) # lower index means higher score
# >>> rank_res: [0, 1]
```
### Open Source License
The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow free commercial usage. To apply for a commercial license, please fill in the application form (English)/申请表（中文）. For other questions or collaborations, please contact internlm@pjlab.org.cn.